arXiv subjects

Yizhou Liu

Publications and source records attributed to Yizhou Liu.

At least 55 records · Page 3Linked to original sources

arXiv2026

Adaptive Reinforcement for Open-ended Medical Reasoning via Semantic-Guided Reward Collapse Mitigation

Reinforcement learning (RL) with rule-based reward functions has recently shown great promise in enhancing the reasoning depth and generalization ability of vision-language models (VLMs), while maintaining computational efficiency. In spite of these advances, its adoption in medical imaging remains limited. Current reinforcement fine-tuning (RFT) efforts in this field mainly focus on closed-ended visual question answering (VQA), restricting their applicability to realistic clinical reasoning. However, open-ended medical VQA better mirrors clinical diagnostic workflows but remains underexplored. Although several studies have attempted to bridge the two formats through semantically guided RL, model-driven semantic rewards often suffer from reward collapse, where responses with distinct semantics yield nearly identical scores. To overcome this limitation, we introduce Adaptive Reinforcement for Medical Reasoning (ARMed), a novel RL framework tailored for open-ended medical VQA. ARMed first injects domain expertise through supervised fine-tuning (SFT) on chain-of-thought annotations, followed by reinforcement optimization using textual correctness and adaptive semantic rewards to refine reasoning consistency and factual accuracy. Extensive experiments on six challenging medical VQA benchmarks demonstrate that ARMed substantially improves both accuracy and generalization. These findings underscore the importance of reward discriminability in medical RL and highlight the potential of adaptive semantic rewards for building robust, clinically reliable multimodal reasoning systems.

cs.CV↗

arXiv2026

Collective Interference of Phonon Spin and Dipole Moment Rotation Induced Circular Dichroism

The classical field description of phonon spin relies on the invariance of a continuous elastic field under infinitesimal rotation. However, a local medium element in the continuous field may contain large numbers of vibrational particles at microscopic level, like for complex lattices with many atoms in a unit cell. We find this causes the phonon spin in real materials no longer a simple sum of each atom rotation, but a collective interference of many atoms, since phonons are phase-coherent vibrational modes across unit cells. We demonstrate the collective interference phonon spin manifested as the dipole moment rotating (DMR) of charge-polarized unit cell, by deriving the infrared circular dichroism (ICD) with phonon-photon interaction in complex lattices. We compare the DMR with the local atom rotation without interference, and exemplify their distinct ICD spectrum in a chiral lattice model and two realistic chiral materials. Detectable ICD measurements are proposed in quartz with Weyl phonon near Gamma point. Our study underlies the important role of collective interference and uncovers a deeper insight of phonon spin in real materials with complex lattices.

cond-mat.mtrl-sci↗

arXiv2026

Halo abundance and clustering in cosmologies with massive and asymmetric neutrinos

Neutrinos are the most abundant fermions in the Universe and influence the formation of large-scale structure through both their non-zero masses and a possible chemical potential which can be described by a single asymmetry parameter. While most previous studies have focused on the impact of the neutrino mass, the role of neutrino asymmetry remains comparatively unexplored. In this work, we investigate how massive neutrinos ($M_ν=0-0.24\,\mathrm{eV}$) with a non-zero asymmetry parameter ($η^{2}=0-0.8$) modify the halo mass function (HMF) and halo bias using cosmological N-body simulations with cosmological parameters consistently refitted to CMB observations. We find that at all redshifts, neutrino mass suppresses the abundance of massive halos, whereas neutrino asymmetry enhances the HMF over a broad mass range. At z=0, the abundance of the most massive halos is reduced by up to ~30% in the largest-mass case ($M_ν=0.24\,\mathrm{eV}$), while neutrino asymmetry ($η^{2}=0.8$) produces a maximum ~5% enhancement. These effects become increasingly pronounced at higher redshifts: by z=4 and z=9, the enhancement induced by neutrino asymmetry reaches ~25% and ~75%, respectively, while the corresponding suppression due to neutrino mass deepens to below ~40% and ~70% of the massless case. For halo bias, we find that halos with masses above $10^{13.4}\,\mathrm{M_\odot}$ exhibit an enhanced large-scale bias due to neutrino mass, reaching up to ~5% at z=0, while neutrino asymmetry reduces the bias by a few percent on linear scales. These trends strengthen with redshift, with the enhancement and suppression growing to ~15% and ~10% at z=2, respectively. Linear bias models provide an adequate, though not exact, description of halo bias in massive-neutrino cosmologies. Our results demonstrate that halo abundance and clustering offer sensitive probes of both neutrino mass and asymmetry.

astro-ph.CO↗

arXiv2026

Beyond Pixel Simulation: Pathology Image Generation via Diagnostic Semantic Tokens and Prototype Control

In computational pathology, understanding and generation have evolved along disparate paths: advanced understanding models already exhibit diagnostic-level competence, whereas generative models largely simulate pixels. Progress remains hindered by three coupled factors: the scarcity of large, high-quality image-text corpora; the lack of precise, fine-grained semantic control, which forces reliance on non-semantic cues; and terminological heterogeneity, where diverse phrasings for the same diagnostic concept impede reliable text conditioning. We introduce UniPath, a semantics-driven pathology image generation framework that leverages mature diagnostic understanding to enable controllable generation. UniPath implements Multi-Stream Control: a Raw-Text stream; a High-Level Semantics stream that uses learnable queries to a frozen pathology MLLM to distill paraphrase-robust Diagnostic Semantic Tokens and to expand prompts into diagnosis-aware attribute bundles; and a Prototype stream that affords component-level morphological control via a prototype bank. On the data front, we curate a 2.65M image-text corpus and a finely annotated, high-quality 68K subset to alleviate data scarcity. For a comprehensive assessment, we establish a four-tier evaluation hierarchy tailored to pathology. Extensive experiments demonstrate UniPath's SOTA performance, including a Patho-FID of 80.9 (51% better than the second-best) and fine-grained semantic control achieving 98.7% of the real-image. The dataset and code can be obtained from https://github.com/Hanminghao/UniPath.

cs.CV↗

arXiv2026

Superposition unifies power-law training dynamics

We investigate the role of feature superposition in the emergence of power-law training dynamics using a teacher-student framework. We first derive an analytic theory for training without superposition, establishing that the power-law training exponent depends on both the input data statistics and channel importance. Remarkably, we discover that a superposition bottleneck induces a transition to a universal power-law exponent of $\sim 1$, independent of data and channel statistics. This one over time training with superposition represents an up to tenfold acceleration compared to the purely sequential learning that takes place in the absence of superposition. Our finding that superposition leads to rapid training with a data-independent power law exponent may have important implications for a wide range of neural networks that employ superposition, including production-scale large language models.

cs.LG↗

arXiv2026

The Blessing of Dimensionality in LLM Fine-tuning: A Variance-Curvature Perspective

Weight-perturbation evolution strategies (ES) can fine-tune billion-parameter language models with surprisingly small populations (e.g., $N\!\approx\!30$), contradicting classical zeroth-order curse-of-dimensionality intuition. We also observe a second seemingly separate phenomenon: under fixed hyperparameters, the stochastic fine-tuning reward often rises, peaks, and then degrades in both ES and GRPO. We argue that both effects reflect a shared geometric property of fine-tuning landscapes: they are low-dimensional in curvature. A small set of high-curvature dimensions dominates improvement, producing (i) heterogeneous time scales that yield rise-then-decay under fixed stochasticity, as captured by a minimal quadratic stochastic-ascent model, and (ii) degenerate improving updates, where many random perturbations share similar components along these directions. Using ES as a geometric probe on fine-tuning reward landscapes of GSM8K, ARC-C, and WinoGrande across Qwen2.5-Instruct models (0.5B--7B), we show that reward-improving perturbations remain empirically accessible with small populations across scales. Together, these results reconcile ES scalability with non-monotonic training dynamics and suggest that high-dimensional fine-tuning may admit a broader class of viable optimization methods than worst-case theory implies.

cs.LG↗

arXiv2026

The 2026 Skyrmionics Roadmap

Magnetic skyrmions and related topological spin textures have emerged as a central topic in condensed-matter physics, combining fundamental significance with potential for transformative applications in spintronics, magnonics, and beyond. Over the past decade, advances in material platforms, imaging techniques, theoretical modeling, and device concepts have established skyrmionics as a rapidly expanding field. At the same time, challenges remain in stabilizing, controlling, and integrating such textures into functional architectures, while novel phenomena such as antiskyrmions, higher-order skyrmions, hopfions, and antiferromagnetic textures arise. The 2026 Skyrmionics Roadmap represents a collective effort of many authors, providing a comprehensive perspective on the current state-of-the-art and the outlook for the coming years. In 33 focused sections, each co-authored by two researchers, we chart progress in theory and modeling, material systems, skyrmion dynamics, and skyrmion technologies. By offering a consolidated vision, this Roadmap aims to guide both fundamental research and application-driven efforts, accelerating the transition of skyrmionics from conceptual breakthroughs toward practical technologies.

cond-mat.mes-hall↗

arXiv2025

Spin and Orbital Angular Momentum Polarization in Thouless Topological Charge Pumping

Quantized charge pumping in one-dimensional chiral wires has been widely studied in the context of topological physics in a (1+1)-dimensional synthetic space, yet the role of orbital and spin degrees of freedom in such topological pumps remains largely unexplored. Here, we examine how the topologically quantized charge pump in insulators generates spin polarizations, and assess whether this mechanism may offer distinct insight into the widely known spin-selective transport in chiral wires-commonly referred to as chirality-induced spin selectivity. We performed time-dependent Schrodinger equations of multi-orbital tight-binding Hamiltonians driven by a circularly polarized electric field. Our main findings are twofold. First, the intrinsic screw-like geometry of the system generates a distinctive winding structure governed by a single control parameter, in contrast to conventional adiabatic pumping mechanisms that require at least two independently modulated parameters, thereby providing a clear interpretation of one-dimensional pumping in terms of the topological structure in a (1+1)-dimensional Brillouin zone. Second, while the energy gap remains open throughout the pumping cycle, the Berry-phase driven real-time dynamics of the charge flow induces a nonequilibrium orbital polarization. Through spin-orbit coupling, this orbital response is partially converted into spin polarization whose direction is determined by the current and chirality. On the analogy between the synthetic (1+1)- and 2-dimensional topological insulators, we suggest that non-trivial spin-orbital dynamics may accompany the anomalous quantum charge Hall states of even-dimensional real materials.

cond-mat.mes-hall↗

arXiv2025

Forging a Dynamic Memory: Retrieval-Guided Continual Learning for Generalist Medical Foundation Models

Multimodal biomedical Vision-Language Models (VLMs) exhibit immense potential in the field of Continual Learning (CL). However, they confront a core dilemma: how to preserve fine-grained intra-modality features while bridging the significant domain gap across different modalities. To address this challenge, we propose a comprehensive framework. Leveraging our 18-million multimodal and comprehensive medical retrieval database derived from PubMed scientific papers, we pioneer the integration of Retrieval-Augmented Generation (RAG) into CL. Specifically, we employ a multi-modal, multi-layer RAG system that provides real-time guidance for model fine-tuning through dynamic, on-demand knowledge retrieval. Building upon this, we introduce a dynamic knowledge distillation framework. This framework precisely resolves the aforementioned core dilemma by dynamically modulating the importance of the parameter space, the granularity of the distilled knowledge, and the data distribution of the reference dataset in accordance with the required level of detail. To thoroughly validate the clinical value of our strategy, we have designed a more rigorous \textbf{M}edical Generalist Task Incremental Learning (MGTIL) benchmark. This benchmark is engineered to simultaneously evaluate the model's capacity for adaptation to significant domain shifts, retention of subtle intra-domain features, and real-time learning of novel and complex medical tasks. Extensive experimental results demonstrate that our proposed method achieves state-of-the-art (SOTA) performance across all metrics. The code is provided in the supplementary materials.

cs.CV↗

arXiv2025

FysicsWorld: A Unified Full-Modality Benchmark for Any-to-Any Understanding, Generation, and Reasoning

Despite rapid progress in multimodal large language models (MLLMs) and emerging omni-modal architectures, current benchmarks remain limited in scope and integration, suffering from incomplete modality coverage, restricted interaction to text-centric outputs, and weak interdependence and complementarity among modalities. To bridge these gaps, we introduce FysicsWorld, the first unified full-modality benchmark that supports bidirectional input-output across image, video, audio, and text, enabling comprehensive any-to-any evaluation across understanding, generation, and reasoning. FysicsWorld encompasses 16 primary tasks and 3,268 curated samples, aggregated from over 40 high-quality sources and covering a rich set of open-domain categories with diverse question types. We also propose the Cross-Modal Complementarity Screening (CMCS) strategy integrated in a systematic data construction framework that produces omni-modal data for spoken interaction and fusion-dependent cross-modal reasoning. Through a comprehensive evaluation of over 30 state-of-the-art baselines, spanning MLLMs, modality-specific models, unified understanding-generation models, and omni-modal language models, FysicsWorld exposes the performance disparities and limitations across models in understanding, generation, and reasoning. Our benchmark establishes a unified foundation and strong baselines for evaluating and advancing next-generation full-modality architectures.

cs.CV↗

arXiv2025

Superposition Yields Robust Neural Scaling

The success of today's large language models (LLMs) depends on the observation that larger models perform better. However, the origin of this neural scaling law, that loss decreases as a power law with model size, remains unclear. We propose that representation superposition, meaning that LLMs represent more features than they have dimensions, can be a key contributor to loss and cause neural scaling. Based on Anthropic's toy model, we use weight decay to control the degree of superposition, allowing us to systematically study how loss scales with model size. When superposition is weak, the loss follows a power law only if data feature frequencies are power-law distributed. In contrast, under strong superposition, the loss generically scales inversely with model dimension across a broad class of frequency distributions, due to geometric overlaps between representation vectors. We confirmed that open-sourced LLMs operate in the strong superposition regime and have loss scaling inversely with model dimension, and that the Chinchilla scaling laws are also consistent with this behavior. Our results identify representation superposition as a central driver of neural scaling laws, providing insights into questions like when neural scaling laws can be improved and when they will break down.

cs.LG↗

arXiv2025

Particle loads for cosmological simulations with equal-mass dark matter and baryonic particles

Traditional cosmological hydrodynamical simulations usually assume equal-numbered but unequal-mass dark matter and baryonic particles, which can lead to spurious collisional heating due to energy equipartition. To avoid such a numerical heating effect, a simulation setup with equal-mass dark matter and baryonic particles, which corresponds to a particle number ratio of $N_{\rm DM}:N_{\rm gas} = Ω_{\rm cdm} / Ω_{\rm b}$, is preferred. However, previous studies have typically used grid-based particle loads to prepare such initial conditions, which can only reach specific values for $N_{\rm DM}:N_{\rm gas}$ due to symmetry requirements. In this study, we propose a method based on the glass approach that can generate two-component particle loads with more general $N_{\rm DM}:N_{\rm gas}$ ratios. The method simultaneously relaxes two Poisson particle distributions by introducing an additional repulsive force between particles of the same component. We show that the final particle load closely follows the expected minimal power spectrum, $P(k) \propto k^{4}$, exhibits good homogeneity and isotropy properties, and remains sufficiently stable under gravitational interactions. Both the dark matter and gas components individually also exhibit uniform and isotropic distributions. We apply our method to two-component cosmological simulations and demonstrate that an equal-mass particle setup effectively mitigates the spurious collisional heating that arises in unequal-mass simulations. Our method can be extended to generate multi-component uniform and isotropic distributions. Our code based on Gadget-2 is available at https://github.com/liaoshong/gadget-2glass .

astro-ph.CO↗

arXiv2025

Phonon Spin Selective One-Way Axial Phonon Transport in Chiral Nanohelix

Selectively exciting and manipulating phonons at nanoscale becomes more and more important but still remains challenging in modern nano-energy control and information sensing. Here, we show that the phonon spin angular momentum provides an extra degree of freedom to achieve versatile manipulation of axial phonons in nanomaterials via coupling to spinful multi-physical fields, such as circularly polarized infrared absorption. In particular, we demonstrate the nanoscale one-way axial phonon excitation and routing in chiral nanomaterials, by converting the photon spin in circularly polarized optical fields into the collective interference phonon spin. As exemplified in the smallest chiral carbon nanotube, we show that the rectification rate can reach nearly 100\%, achieving an ideal one-way phonon router, which is verified by molecular dynamics simulations. Our results shed new light on the flexible phonon manipulation via phonon spin degree of freedom, paving the way for future spin phononics.

cond-mat.mes-hall↗

arXiv2025

Trae Agent: An LLM-based Agent for Software Engineering with Test-time Scaling

Software issue resolution is a critical challenge in software engineering and has garnered increasing attention in recent years. With the rapid advancement of large language models (LLMs), substantial progress has been made in addressing real-world software engineering tasks. Recent studies have introduced ensemble reasoning techniques to enhance the performance of LLM-based issue resolution. However, existing prompting-based methods still face limitations in effectively exploring large ensemble spaces and lack the capacity for repository-level understanding, both of which constrain their overall effectiveness. In this paper, we propose Trae Agent, the first agent-based ensemble reasoning approach for repository-level issue resolution. Trae Agent formulates our goal as an optimal solution search problem and addresses two key challenges, i.e., large ensemble spaces and repository-level understanding, through modular agents for generation, pruning, and selection. We conduct extensive experiments using three leading LLMs on the widely-adopted SWE-bench benchmark, comparing Trae Agent against four state-of-the-art ensemble reasoning techniques. Experimental results demonstrate that Trae Agent consistently achieves superior performance, with an average improvement of 10.22% over all baselines in terms of Pass@1. Trae Agent has achieved first place on the SWE-bench Verified leaderboard, with a notable Pass@1 score of 75.20%. We are pleased to release Trae Agent as an open-source project to support the research community, with all resources available at https://github.com/bytedance/trae-agent.

cs.SE↗

arXiv2025

Neural Thermodynamic Laws for Large Language Model Training

Beyond neural scaling laws, little is known about the laws underlying large language models (LLMs). We introduce Neural Thermodynamic Laws (NTL) -- a new framework that offers fresh insights into LLM training dynamics. On the theoretical side, we demonstrate that key thermodynamic quantities (e.g., temperature, entropy, heat capacity, thermal conduction) and classical thermodynamic principles (e.g., the three laws of thermodynamics and the equipartition theorem) naturally emerge under river-valley loss landscape assumptions. On the practical side, this scientific perspective yields intuitive guidelines for designing learning rate schedules.

cs.LG↗

arXiv2025

VLM-based Prompts as the Optimal Assistant for Unpaired Histopathology Virtual Staining

In histopathology, tissue sections are typically stained using common H&E staining or special stains (MAS, PAS, PASM, etc.) to clearly visualize specific tissue structures. The rapid advancement of deep learning offers an effective solution for generating virtually stained images, significantly reducing the time and labor costs associated with traditional histochemical staining. However, a new challenge arises in separating the fundamental visual characteristics of tissue sections from the visual differences induced by staining agents. Additionally, virtual staining often overlooks essential pathological knowledge and the physical properties of staining, resulting in only style-level transfer. To address these issues, we introduce, for the first time in virtual staining tasks, a pathological vision-language large model (VLM) as an auxiliary tool. We integrate contrastive learnable prompts, foundational concept anchors for tissue sections, and staining-specific concept anchors to leverage the extensive knowledge of the pathological VLM. This approach is designed to describe, frame, and enhance the direction of virtual staining. Furthermore, we have developed a data augmentation method based on the constraints of the VLM. This method utilizes the VLM's powerful image interpretation capabilities to further integrate image style and structural information, proving beneficial in high-precision pathological diagnostics. Extensive evaluations on publicly available multi-domain unpaired staining datasets demonstrate that our method can generate highly realistic images and enhance the accuracy of downstream tasks, such as glomerular detection and segmentation. Our code is available at: https://github.com/CZZZZZZZZZZZZZZZZZ/VPGAN-HARBOR

eess.IV↗

arXiv2025

Construction of Hopfion Crystals

Hopfions, three-dimensional topological solitons characterized by nontrivial Hopf indices, represent a fundamental class of field configurations that emerge across diverse areas of physics. Despite extensive studies of isolated hopfions, a framework for constructing spatially ordered arrays of hopfions, i.e., hopfion crystals, has been lacking. Here, we present a systematic approach for generating hopfion crystals with cubic symmetry by combining the Hopf map with rational mapping techniques. By superposing helical waves in $\mathbb{R}^4$, we construct hopfion crystals with tunable Hopf indices and controllable topology. We demonstrate simple cubic, facecentered cubic, and body-centered cubic hopfion crystals, and extend our framework to create crystals of more complex topological structures, including axially symmetric tori, torus links, and torus knots with higher Hopf indices. Our results provide a foundation for searching hopfions in real materials and studying their collective phenomena.

cond-mat.str-el↗

arXiv2025

The mass accretion history of dark matter haloes down to Earth mass

We take advantage of the unprecedented dynamical range provided by the "Cosmic-Zoom" project to study the mass accretion history (MAH) of present-day dark matter haloes over the entire mass range present in the $Λ$CDM paradigm when the dark matter is made of weakly interacting massive particles of mass $100\ \mathrm{GeV}$. In particular, we complement previous studies by exploring the MAHs of haloes with mass from $10^8\ h^{-1}\mathrm{M_{\odot}}$ down to Earth mass, $10^{-6}\ h^{-1}\mathrm{M_{\odot}}$. The formation redshift of low-mass haloes anti-correlates weakly with mass, peaking at $z=3$ for haloes of mass $10^{-4}\ h^{-1}\mathrm{M_{\odot}}$. Even lower masses are affected by the free-streaming cutoff in the primordial spectrum of density fluctuations and form at lower redshift. We compare MAHs in our simulations with predictions from two analytical models based on the extended Press-Schechter theory (EPS), and three empirical models derived by fitting and extrapolating either results from cosmological $N$-body simulations or Monte Carlo realizations of halo growth. All models fit our simulations reasonably well over the mass range for which they were calibrated. While the empirical models match better for more massive haloes, $M>10^{10}\ h^{-1}\mathrm{M_{\odot}}$, the analytical models do better when extrapolated down to Earth mass. At the higher masses, we explore the correlation between local environment density and MAH, finding that biases are relatively weak, with typical MAHs for haloes in extremely low-density and in typical regions differing by less than $20$ percent at high redshift. If this result can be extrapolated to lower halo masses, we conclude that EPS theory predicts the hierarchical build-up of dark matter haloes quite well over the entire halo mass range.

astro-ph.GA↗